Soft Computing

1
Soft Computing

• Lecture 16
• Spiking neural networks

2
Phenomenology of spike generation
threshold -gt Spike
i
Threshold ?Spike emission (Action potential)
3
The problem of neural coding temporal codes
Time to first spike after input
correlations
Phase with respect to oscillation
4
Rank Order Coding
One possibility takes advantage of the fact that
a neuron can be thought of as an analog-delay
convertor. It acts somewhat like a capacitance
which is progressively charged by an input until
it reaches a threshold, at which point it
generates an output pulse the action potential
or spike. Such neurons will naturally fire
earliest when the input is strong, and will take
progressively longer to fire when the input is
weaker. In this way, the time at which a neuron
fires (its response latency) can be used to code
the intensity of the stimulus.
However, this sort of code requires knowledge of
when the stimulation started, information which
is not generally available in the case of the
biological visual system. There is, however, a
way round this. Consider what happens when
several neurons are used in parallel. In this
case, even without knowing the precise moment at
which the stimulus came on, information can be
obtained by looking at the order in which the
neurones fire
The order of firing of a group of neurons is
potentially a very rich source of information
about the input pattern
5
SpikeNet

• Is developed for control systems
• Features of neuron
• Feedback from output to inputs for updating of
  weights
• At summation of signals take account of frequency
  of signal
• Activation function is threshold function
• Award is used for learning of neuron
• Inputs is discrete
• One-layer network from this neurons is able to
  execute functions which available only for multi
  layer recurrent networks

6
The neural network architecture SEQAINET
(SEQence association and Integration NETwork).
7
FPGA hardware spike neural network for robots
8
FPGA hardware spike neural network for robots
(2)